It's a losing battle trying to get every technology and web service to put in all the right aria tags - imagine imgur for example - there is no way you could make those all those animated gifs and mspaint memes properly accessible - instead, a neural network should look at the screen like a human does, and describe it.
The screen reader should say "There is a gif of a squirrel falling off a shelf, and it has a share and an upload button"